Mask Wearing Detection And Embedded Implementation Based On YOLOv4 Block Weight Pruning

Mask wearing detection is a kind of target detection technology.Wearing masks is a requirement for regular epidemic prevention and control.The main purpose of this study is to utilize the embedded devices to realize front-end detection of mask wearing in communities,shopping malls,railway stations,airports and other places with large flow of people.This paper proposed a lightweight target detection network framework YOLOK210,based on YOLOv4 to reduce the equipment volume,cost and deployment difficulty.Firstly,a block-weight pruning algorithm suitable for arbitrary kernel size is used to compress the network model.Then the data set is preprocessed and the non-maximum suppression algorithm is used to reduce the positioning error.Finally,the detection network is transplanted to the embedded platform,that based on K210 chip.The purpose of real-time detection is achieved.The main work and innovation are as follows:(1)Aiming at the problem of large model in the target detection network based on deep learning,the block weight pruning method is adopted to compress the model size on YOLOv4 and reduce the number of parameters in the detection network.The purpose of improving the detection speed without sacrificing the detection accuracy is achieved,and the whole network frame is more suitable for the embedded implementation of mask wearing detection.(2)Aiming at the problems of high cost and high power consumption of the existing embedded platform based on deep learning,the hardware design and construction of the embedded platform are completed by using the domestic chip K210,and the overall framework is further optimized by using the method of compiler optimization.The mask wearing detection is realized on embedded device,and the ideal detection speed and accuracy are achieved.(3)From the perspective of practical engineering application,the idea of promoting the implementation of deep learning technology runs through the whole project.In the whole implementation process,the complexity,cost and application scenarios are considered comprehensively.In line with the actual application requirements,it can complete the mask wearing detection and have certain practical value.The experiment shows that the mask wearing detection algorithm applied in this study can correctly judge whether the subjects wear masks or not,meanwhile maintain a high network compression ratio and recognition speed with less loss of accuracy.On RTX2080,after 14 times of compression,mean average precision(m AP)can reach 85.2%and frames per second(FPS)can reach 30.Transplanted to K210,m AP is 64.8% and FPS is 15.4.It can achieve the basic requirements of real-time detection,and complete mask wearing detection and embedded implementation.